Who Secures the Agents?
In this episode of Application Security Weekly, host Mike Shima and co-host John Canella sit down with Dan Moore (Senior Director of CIAM Strategy and Identity Standards at FusionAuth) to dig into one of the most pressing — and underexplored — questions in AI security: how do you bring real authentication and authorization to the world of agents and MCPs?
The answer isn't just "use OAuth." It's about understanding what granular access control actually looks like when your software can exhaust every possibility at machine speed.
Inside the Episode:
- Don't Reinvent the Wheel: Dan draws direct parallels between the API security lessons of the early 2010s — discovery, gateways, centralized policy — and the emerging MCP ecosystem. The problems aren't new; the actors are.
- The Authorization Gap: MCP leverages OAuth 2.1 for coarse-grained remote access, but the fine-grained authorization behind MCP servers is left entirely to implementers. Dan's Google Drive example illustrates the real-world friction: restricting an agent to a single folder required creating a service account outside of MCP entirely.
- RBAC, ABAC, and ReBAC — Oh My: A practical walkthrough of role-based, attribute-based, and relationship-based access control models, and why choosing the right one for your agent architecture matters more than ever when non-deterministic entities are probing every guardrail.
- Centralize or Suffer: The case for externalizing authorization logic rather than scattering it throughout code — better visibility, easier auditing, testable policies, and consistency across multiple APIs and applications.
- The Consent Fatigue Anti-Pattern: If agents ask users for permission five times a minute, everyone clicks "yes." The real answer is building intelligent, developer-side guardrails rather than relying on human approval at runtime.
